Skip to content

Back to the catalog

AutoJack Attack Lets One Web Page Hijack AI Agent for Host Code Execution
thehackernews.comarticle

AutoJack Attack Lets One Web Page Hijack AI Agent for Host Code Execution

Microsoft researchers detailed AutoJack, an exploit chain in AutoGen Studio, Microsoft Research's multi-agent framework prototyping UI, where a browsing AI agent that loads an attacker-controlled webpage inherits localhost trust, letting the page's JavaScript reach an unauthenticated MCP WebSocket handler that runs an attacker-chosen command with no allowlist. The Hacker News independently confirmed the vulnerable handler shipped only in two PyPI pre-releases, 0.4.3.dev1 and 0.4.3.dev2, not the stable 0.4.2.2 release, and it remains unpatched on PyPI even though maintainers hardened GitHub main at commit b047730 after Microsoft's MSRC report. The piece frames this as a generalizable pattern for agent frameworks, localhost checks treated as auth, unauthenticated local control planes, and agents that browse untrusted content, also seen in the ChatGPhish case and Microsoft's Semantic Kernel RCE research.

TL;DR A browsing AI agent that opens an attacker's webpage can trigger unauthenticated RCE via AutoGen Studio's unpatched MCP WebSocket handler, which shipped only in two un-yanked PyPI pre-releases.
Document

What it is

AutoJack is a research exploit chain in AutoGen Studio (Microsoft Research's AutoGen multi-agent prototyping UI). The vulnerable MCP WebSocket handler shipped only in PyPI pre-releases 0.4.3.dev1 and 0.4.3.dev2 — the stable 0.4.2.2 has no MCP route at all. Both dev builds remain on PyPI, unyanked, and no released build carries the fix; the patch exists only in GitHub main at commit b047730 (PR #7362).

What to do

The summary establishes the chain and the packaging boundary but gives no remediation. The page's concrete guidance:

  • Plain pip install autogenstudio gives 0.4.2.2 (no MCP route) — not affected. pip won't pull the dev builds unless you pass --pre or pin the version.
  • If you installed a pre-release, there is no patched PyPI build to upgrade to. Pull from GitHub main at or after commit b047730 — the article calls this "the real fix."
  • Until a release ships, break the shared-localhost precondition: do not run AutoGen Studio on the same machine as a browsing or code-execution agent that touches untrusted content. If they must co-locate, isolate in separate containers/VMs and run AutoGen Studio under a low-privilege account.

The fix mechanism

The hardened handler no longer reads the command from the URL. Request parameters are stored server-side behind a one-time session ID, and unknown IDs are refused. MCP routes now pass through the normal authentication middleware rather than being skipped.

flowchart TD
    A[Agent loads attacker page] --> B{localhost check}
    B -->|agent IS localhost| C{auth middleware}
    C -->|MCP paths skipped, handler never verifies| D[command from request param]
    D -->|no allowlist| E[RCE under AutoGen Studio account]
    C -.->|b047730: MCP routes now authenticated| X[blocked]
    D -.->|b047730: one-time session ID, cmd not in URL| X

Also

The page names the Semantic Kernel RCE the summary references only generically: CVE-2026-26030 and CVE-2026-25592. Microsoft reports no exploitation in the wild; the PoC merely pops calc.exe.

Organisations named here

  • AutoJack
  • ChatGPhish

CVEs named here

  • CVE-2026-26030
  • CVE-2026-25592

Technologies named here

  • AutoGen - Microsoft Research's multi-agent framework
  • Microsoft Semantic Kernel
  • ChatGPT - referenced re
  • Model Context Protocol
  • AutoGen Studio
  • PyPI
  • pip
  • Semantic Kernel

Techniques named here

  • Exploitation of a localhost trust check
  • Authentication bypass
  • Command injection/unsanitized command execution